Localized Kernel Methods for Signal Processing

By · 2025

Read Localized Kernel Methods for Signal Processing by By (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-04978.txt.

Introduction

Sippanon Kitimoon presents in a 2025 dissertation two signal processing methods built on localized kernels, math tools that focus on small slices of a signal. Both methods recover hidden parameters from noisy data without heavy assumptions. The first finds the pitches and strengths inside a mix of pure tones that fade or grow.

That method beat two classic tools when noise ran high, and a projection trick let it handle multi-part signals with few samples. The second method splits apart sliding tones, called chirps, even when they cross or break off, at noise levels far below the signal. Both share fast Fourier-based code and need no sparsity guesses.

Catalog record

CREATORBy
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE128 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-04978

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Localized Kernel Methods for Signal Processing By Abstract © Copyright Sippanon Kitimoon, 2025. All rights reserved Approval of the Dissertation Committee This dissertation has been duly read, reviewed, and critiqued by the Committee listed below, which hereby approves the manuscript of Sippanon Kitimoon as fulfilling the scope and quality requirements for meriting the degree of Doctor of Philosophy in Mathematics. Hrushikesh Mhaskar, Chair Claremont Graduate University Distinguished Research Professor of Mathematics Qidi Peng Claremont Graduate University Research Associate Professor of Mathematics Allon Percus Claremont Graduate University Joseph H. Pengilly Professor of Mathematics Abstract Localized Kernel Methods for Signal Processing By Sippanon Kitimoon Claremont Graduate University: 2025 This dissertation presents two signal processing methods using specially designed localized kernels for parameter recovery under noisy condition. The first method addresses the estimation of frequencies and amplitudes in multidimensional exponential models. It utilizes localized trigonometric polynomial kernels to detect the multivariate frequencies, followed by a more detailed parameter estimation. We compare our method with MUSIC and ESPRIT, which are classical subspace-based algorithms widely used for estimating the parameters of exponential signals. In the univariate case, the method outperforms MUSIC and ESPRIT under low signal-to-noise ratios. For the multivariate case, we develop a coordinate-wise projection and registration approach that achieves high recovery accuracy using significantly fewer samples than other methods. The second method focuses on separating linear chirp components from time-localized signal segments. A variant of the Signal Separation Operator (SSO) is constructed using a localized kernel. Instantaneous frequency estimates are obtained via FFT-based filtering, then clustered and fitted with piecewise linear regression. The method operates without prior knowledge of the number of components and is shown to recover intersecting and discontinuous chirps at SNR levels as low as -30 dB. Both methods share an idea based on localized kernels and efficient FFT-based implementation, and neither requires subspace decomposition or sparsity regularization. Experimental results confirm the robustness and tractability of the proposed approaches across a range of simulated data conditions. Potential extensions include application to nonlinear chirps, adaptive kernel design, and signal classification using extracted features. Acknowledgements I would like to express my deepest gratitude to my Ph.D. advisor, Prof. Hrushikesh Mhaskar, for his unwavering support, insightful guidance, and patient mentorship throughout every stage of this work. His mathematical rigor, clarity of thought, and encouragement have been essential to my development as a researcher. Beyond his academic mentorship, I am especially thankful for the parental care and support he provided during my Ph.D. journey, for which I am sincerely grateful. I am also grateful to the other members of my dissertation committee, Prof. Allon Percus and Prof. Qidi Peng, for agreeing to serve on my dissertation committee. I greatly appreciate their time, dedication, and encouragement during my doctoral journey. I would like to thank Dr. Raghu G. Raj and Dr. Eric Mason for their essential contributions to this work. Dr. Raj provided helpful direction and recommendations during the development of the methods presented in multivariate the exponential analysis portion of the dissertation. Dr. Mason generated simulation data and offered key guidance and suggestions in the formulation and implementation of the chirp signal separation approach. Their involvement improved both the clarity and practical relevance of the results. Finally, I am thankful to all those who have supported and encouraged me in both work and recreational environment. I wish them all the best. List of important symbols \zeta(S) Riemann zeta function \delta(t) Dirac delta function \delta_{y} Dirac delta supported at y \Delta_{j} Independent vectors in {\mathbb{R}}^{q} P(z) Prony polynomial c_{j} Coefficients of the Prony polynomial \|\cdot\|_{2} L2 norm \mathbf{H}_{L,N-L+1} Rectangular Hankel matrix e_{L}(\varphi) The column space of \mathbf{H}_{L,N-L+1} \mathcal{S}_{L} The signal space \mathcal{N}_{L} The noise space \widetilde{W}_{N-L+1,K} The first K right singular vectors of \mathbf{H}_{L,N-L+1} V_{N,K}(z) The Vandermonde matrix \epsilon Random variable (noise) \eta Minimal separation among unitless frequencies \hat{\mu} Fourier coefficient of \mu \hbar_{n} Normalizing constant, Eqn. ( 2.1.1 ) \lambda_{k} Univariate unitless frequencies \Lambda_{j,k} Estimated IF for chirp j at t_{k} \mathbf{M} Sum of absolute values of amplitudes \mathcal{T}_{n,R} Signal separation operator \mathbb{G} Support of the thresholded power spectrum \mathbb{G}_{\ell} Cluster definded in Theorem 2.3.1 \mathfrak{m} Minimum absolute value of coefficients \omega_{j,k} , \gamma_{j,k} , a_{j,k} Chirp parameters for chirp j in I_{k} \mu Discretely supported measure/distribution on \mathbb{T} \Phi_{n} Localized kernel of degree n {\mathbb{R}} Set of real numbers \sigma_{n} Reconstruction operator, Section 1.1 \tilde{\mu} Fourier coefficients of \mu plus an additive noise \mathbb{T} Quotient space of real line modulo 2\pi \mathbf{w}_{k} Multivariate frequencies/points {\mathbb{Z}} Set of integers A_{k} Coefficients (complex amplitudes) C Eqn. ( 2.3.6 ) E_{n} Reconstruction operator used with noise alone Eqn. ( 2.3.3 ) H Smooth low pass filter i \sqrt{-1} K Number of exponential signals L Constant in Eqn. ( 2.1.10 ) M Sum of absolute values of coefficients q Dimension of the observations S Localization power Eqn. ( 2.1.10 ) \tau Threshold for SSO \theta , \gamma , B , d Chirp parameters A_{j}(t) , \phi_{j}(t) Instantaneous complex amplitudes and phases B_{\mbox{\scriptsize{rec}}} Receiver bandwidth I_{k} Time interval for snippet [t_{k}-\Delta,t_{k}+\Delta] R Sampling frequency t_{k} Center of the time interval for snippet \nu Parameter (variance) of sub-Gaussian variable PRI Pulse repetition interval Chapter 1 Introduction This dissertation focuses on two core applications in signal processing: • Robust and Tractable Multidimensional Exponential Analysis • Localized Kernel Method for Separation of Linear Chirps The foundation of both applications is built upon the constant parameter problem formally defined in Section 1.1.3 , where the goal is to recover unknown parameters of signals composed of complex exponentials or chirp components from noisy observations. By designing and utilizing localized kernel methods, we develop robust algorithms that perform well in the presence of sub-Gaussian noise. Since the two applications addressed in this dissertation focus on solving distinct signal processing problems, the literature reviews for each application are presented separately in their respective chapters. In this chapter, we will focus on defining the constant parameter problem that underlies both applications and introducing the state-of-the-art methods used to solve the problem. 1.1 Background and problem statement We begin with a brief review of the sub-Gaussian noise model in Section 1.1.1 , followed by a definition of the signal-to-noise ratio used throughout this work in Section 1.1.2 . Section 1.1.3 concludes with a formal statement of the constant parameter problem, which serves as the foundational model for the applications developed in later chapters. 1.1.1 Sub-Gaussian noise In this section, we review certain properties of sub-Gaussian random variables. Intuitively, a sub-Gaussian random variable is one whose probability tails decay at least as fast as those of a Gaussian random variable; i.e. extreme deviations are exponentially unlikely and no heavier than in the normal distribution. The material in this section is based on [ 1 , Section 2.3] , with a slight change of notation. A mean zero real valued random variable X is called sub-Gaussian with parameter \nu ( X\in\mathcal{G}(\nu) ) if \log\mathbb{E}(\exp(tX))\leq(t\nu)^{2}/2. (1.1.1) Examples include Gaussian variables and all bounded random variables. For a sub-Gaussian variable, it is shown in [ 1 , Section 2.3] that condition ( 1.1.1 ) implies the tail bound \mathsf{Prob}(|X|>t)\leq 2\exp(-t^{2}/(2\nu^{2})). (1.1.2) Conversely, by [ 1 , Section 2.3, Theorem 2.1] , the tail bound in ( 1.1.2 ) implies X\in\mathcal{G}(4\nu) .

Read the complete text